import pandas as pd

initial_df = pd.read_csv('../../data/ge.us.txt')
initial_df

	Day	Open	High	Low	Close	Volume	OpenInt
0	1962-01-02	0.6277	0.6362	0.6201	0.6201	2575579	0
1	1962-01-03	0.6201	0.6201	0.6122	0.6201	1764749	0
2	1962-01-04	0.6201	0.6201	0.6037	0.6122	2194010	0
3	1962-01-05	0.6122	0.6122	0.5798	0.5957	3255244	0
4	1962-01-08	0.5957	0.5957	0.5716	0.5957	3696430	0
...	...	...	...	...	...	...	...
14053	2017-11-06	20.5200	20.5300	20.0800	20.1300	60641787	0
14054	2017-11-07	20.1700	20.2500	20.1200	20.2100	41622851	0
14055	2017-11-08	20.2100	20.3200	20.0700	20.1200	39672190	0
14056	2017-11-09	20.0400	20.0710	19.8500	19.9900	50831779	0
14057	2017-11-10	19.9800	20.6800	19.9000	20.4900	100698474	0

14058 rows × 7 columns

# Keep only columns Day and Close
initial_df = initial_df[["Day", "Close"]]
initial_df

	Day	Close
0	1962-01-02	0.6201
1	1962-01-03	0.6201
2	1962-01-04	0.6122
3	1962-01-05	0.5957
4	1962-01-08	0.5957
...	...	...
14053	2017-11-06	20.1300
14054	2017-11-07	20.2100
14055	2017-11-08	20.1200
14056	2017-11-09	19.9900
14057	2017-11-10	20.4900

14058 rows × 2 columns

import plotly.express as px

fig = px.line(initial_df, x = "Day", y = "Close", title = "GE Stock")
fig.update_layout(yaxis_title="Close Value")
fig

Normalize the data

from sklearn.preprocessing import StandardScaler
import numpy as np

scaler = StandardScaler()
df = initial_df.copy()

df['Close'] = np.sqrt(df['Close'])
df[['Close']] = scaler.fit_transform(df[['Close']])

#keep centers for denormalization later
center_history = scaler.mean_[0]
scale_history = scaler.scale_[0]

print(center_history)
print(scale_history)

2.677744836440367
1.8340784204156466

ge_trn  = df[0:12000] #training
ge_val  = df[12000:13000] #validation
ge_test = df[13000:14053] #test for deletion

n_inputs = 10 #number of inputs in the RNN e.g. 1st it. use first 10 days to predict the 11th
n_predictions = 1
batch_size = 10

Functions

#tranforms data into the  windows of (10+1) 14,000x11
def build_windowed_matrix(data, timesteps):
    windowed_matrix = np.vstack(
        [data[x : (x + timesteps)] for x in range(len(data) - timesteps + 1)])
    return windowed_matrix

#to do it 14kx11x1 since this is required by keras. f it was multivariate (n) it should be  14kx11xn
def reshape_3D(df):
    return np.reshape(df, (df.shape[0], df.shape[1], 1))

#for each row gets the the x's (10 in number) 
def get_x(mtx, n_inputs, batch_size):
    mtx = mtx[:, 0:n_inputs] 
    mtx = mtx[0:(mtx.shape[0] // batch_size * batch_size), :]
    reshaped_mtx = reshape_3D(mtx)
    return reshaped_mtx

#for each row gets the the y (11th element) + put them in 3D
def get_y(mtx, n_inputs, n_predictions, batch_size):
    mtx = mtx[:, n_inputs : (n_inputs + n_predictions)]
    if (n_inputs + n_predictions) < batch_size:
         mtx = mtx[0:(mtx.shape[0] // batch_size * batch_size),:]
    if n_predictions == 1:
        mtx = np.reshape(mtx, (len(mtx), 1))
    
    reshaped_mtx = reshape_3D(mtx)
    return reshaped_mtx

Extract ‘Close’ Values

trn  = ge_trn['Close'].values.flatten()
val  = ge_val['Close'].values.flatten()
test = ge_test['Close'].values.flatten()

Build matrices

trn_mtx  = build_windowed_matrix(trn, n_inputs+n_predictions)
val_mtx  = build_windowed_matrix(val, n_inputs+n_predictions)
test_mtx = build_windowed_matrix(test, n_inputs+n_predictions)

X_train = get_x(trn_mtx, n_inputs, batch_size)
Y_train = get_y(trn_mtx, n_inputs, n_predictions, batch_size)
X_val   = get_x(val_mtx, n_inputs, batch_size)
Y_val   = get_y(val_mtx, n_inputs, n_predictions, batch_size)
X_test  = get_x(test_mtx, n_inputs, batch_size)
Y_test  = get_y(test_mtx, n_inputs, n_predictions, batch_size)

Build Model

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.callbacks import EarlyStopping

model = keras.Sequential()
model.add(layers.LSTM(units = 64, batch_input_shape=(batch_size, n_inputs, n_predictions)))
model.add(layers.Dense(units = 1))

model.compile(
    loss=keras.losses.logcosh,
    optimizer="sgd",
    metrics=["MeanSquaredError"],
)
model.summary()
callbacks = [
    EarlyStopping(patience=10)
]

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 lstm (LSTM)                 (10, 64)                  16896     
                                                                 
 dense (Dense)               (10, 1)                   65        
                                                                 
=================================================================
Total params: 16,961
Trainable params: 16,961
Non-trainable params: 0
_________________________________________________________________

history = model.fit(x = X_train, y = Y_train, validation_data = [X_val, Y_val], batch_size = batch_size, epochs = 100,
                    callbacks = callbacks)

Epoch 1/100
1199/1199 [==============================] - 8s 5ms/step - loss: 0.0140 - mean_squared_error: 0.0320 - val_loss: 0.0030 - val_mean_squared_error: 0.0060
Epoch 2/100
1199/1199 [==============================] - 6s 5ms/step - loss: 0.0011 - mean_squared_error: 0.0022 - val_loss: 0.0022 - val_mean_squared_error: 0.0044
Epoch 3/100
1199/1199 [==============================] - 6s 5ms/step - loss: 8.1171e-04 - mean_squared_error: 0.0016 - val_loss: 9.5353e-04 - val_mean_squared_error: 0.0019
Epoch 4/100
1199/1199 [==============================] - 6s 5ms/step - loss: 6.8334e-04 - mean_squared_error: 0.0014 - val_loss: 8.8800e-04 - val_mean_squared_error: 0.0018
Epoch 5/100
1199/1199 [==============================] - 6s 5ms/step - loss: 6.0549e-04 - mean_squared_error: 0.0012 - val_loss: 8.2734e-04 - val_mean_squared_error: 0.0017
Epoch 6/100
1199/1199 [==============================] - 5s 5ms/step - loss: 5.5050e-04 - mean_squared_error: 0.0011 - val_loss: 7.5139e-04 - val_mean_squared_error: 0.0015
Epoch 7/100
1199/1199 [==============================] - 6s 5ms/step - loss: 5.1089e-04 - mean_squared_error: 0.0010 - val_loss: 6.4755e-04 - val_mean_squared_error: 0.0013
Epoch 8/100
1199/1199 [==============================] - 6s 5ms/step - loss: 4.7804e-04 - mean_squared_error: 9.5801e-04 - val_loss: 6.0350e-04 - val_mean_squared_error: 0.0012
Epoch 9/100
1199/1199 [==============================] - 5s 5ms/step - loss: 4.5356e-04 - mean_squared_error: 9.0888e-04 - val_loss: 5.8911e-04 - val_mean_squared_error: 0.0012
Epoch 10/100
1199/1199 [==============================] - 6s 5ms/step - loss: 4.3285e-04 - mean_squared_error: 8.6734e-04 - val_loss: 5.0282e-04 - val_mean_squared_error: 0.0010
Epoch 11/100
1199/1199 [==============================] - 6s 5ms/step - loss: 4.1460e-04 - mean_squared_error: 8.3073e-04 - val_loss: 4.1915e-04 - val_mean_squared_error: 8.3872e-04
Epoch 12/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.9963e-04 - mean_squared_error: 8.0069e-04 - val_loss: 5.7115e-04 - val_mean_squared_error: 0.0011
Epoch 13/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.8656e-04 - mean_squared_error: 7.7449e-04 - val_loss: 4.6898e-04 - val_mean_squared_error: 9.3845e-04
Epoch 14/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.7532e-04 - mean_squared_error: 7.5195e-04 - val_loss: 4.3978e-04 - val_mean_squared_error: 8.8001e-04
Epoch 15/100
1199/1199 [==============================] - 7s 6ms/step - loss: 3.6518e-04 - mean_squared_error: 7.3161e-04 - val_loss: 5.2003e-04 - val_mean_squared_error: 0.0010
Epoch 16/100
1199/1199 [==============================] - 7s 5ms/step - loss: 3.5610e-04 - mean_squared_error: 7.1341e-04 - val_loss: 3.8942e-04 - val_mean_squared_error: 7.7921e-04
Epoch 17/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.4812e-04 - mean_squared_error: 6.9741e-04 - val_loss: 4.9455e-04 - val_mean_squared_error: 9.8964e-04
Epoch 18/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.4025e-04 - mean_squared_error: 6.8162e-04 - val_loss: 3.7894e-04 - val_mean_squared_error: 7.5823e-04
Epoch 19/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.3395e-04 - mean_squared_error: 6.6900e-04 - val_loss: 3.7507e-04 - val_mean_squared_error: 7.5050e-04
Epoch 20/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.2855e-04 - mean_squared_error: 6.5818e-04 - val_loss: 3.8867e-04 - val_mean_squared_error: 7.7772e-04
Epoch 21/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.2289e-04 - mean_squared_error: 6.4683e-04 - val_loss: 4.2802e-04 - val_mean_squared_error: 8.5647e-04
Epoch 22/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.1773e-04 - mean_squared_error: 6.3648e-04 - val_loss: 4.1484e-04 - val_mean_squared_error: 8.3010e-04
Epoch 23/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.1265e-04 - mean_squared_error: 6.2631e-04 - val_loss: 3.4836e-04 - val_mean_squared_error: 6.9703e-04
Epoch 24/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.0905e-04 - mean_squared_error: 6.1908e-04 - val_loss: 4.0774e-04 - val_mean_squared_error: 8.1587e-04
Epoch 25/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.0492e-04 - mean_squared_error: 6.1082e-04 - val_loss: 3.7052e-04 - val_mean_squared_error: 7.4140e-04
Epoch 26/100
1199/1199 [==============================] - 6s 5ms/step - loss: 3.0046e-04 - mean_squared_error: 6.0184e-04 - val_loss: 3.6940e-04 - val_mean_squared_error: 7.3915e-04
Epoch 27/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.9738e-04 - mean_squared_error: 5.9569e-04 - val_loss: 3.8177e-04 - val_mean_squared_error: 7.6391e-04
Epoch 28/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.9372e-04 - mean_squared_error: 5.8835e-04 - val_loss: 3.5979e-04 - val_mean_squared_error: 7.1991e-04
Epoch 29/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.9065e-04 - mean_squared_error: 5.8220e-04 - val_loss: 3.7945e-04 - val_mean_squared_error: 7.5926e-04
Epoch 30/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.8758e-04 - mean_squared_error: 5.7605e-04 - val_loss: 3.8973e-04 - val_mean_squared_error: 7.7983e-04
Epoch 31/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.8526e-04 - mean_squared_error: 5.7140e-04 - val_loss: 3.4501e-04 - val_mean_squared_error: 6.9032e-04
Epoch 32/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.8166e-04 - mean_squared_error: 5.6418e-04 - val_loss: 3.2166e-04 - val_mean_squared_error: 6.4360e-04
Epoch 33/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.7964e-04 - mean_squared_error: 5.6015e-04 - val_loss: 3.5519e-04 - val_mean_squared_error: 7.1069e-04
Epoch 34/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.7709e-04 - mean_squared_error: 5.5502e-04 - val_loss: 3.7536e-04 - val_mean_squared_error: 7.5108e-04
Epoch 35/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.7478e-04 - mean_squared_error: 5.5040e-04 - val_loss: 3.5360e-04 - val_mean_squared_error: 7.0752e-04
Epoch 36/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.7267e-04 - mean_squared_error: 5.4616e-04 - val_loss: 3.5248e-04 - val_mean_squared_error: 7.0528e-04
Epoch 37/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.7023e-04 - mean_squared_error: 5.4128e-04 - val_loss: 3.8327e-04 - val_mean_squared_error: 7.6690e-04
Epoch 38/100
1199/1199 [==============================] - 7s 6ms/step - loss: 2.6846e-04 - mean_squared_error: 5.3775e-04 - val_loss: 3.1528e-04 - val_mean_squared_error: 6.3082e-04
Epoch 39/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.6640e-04 - mean_squared_error: 5.3359e-04 - val_loss: 3.1396e-04 - val_mean_squared_error: 6.2818e-04
Epoch 40/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.6495e-04 - mean_squared_error: 5.3071e-04 - val_loss: 3.2290e-04 - val_mean_squared_error: 6.4609e-04
Epoch 41/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.6245e-04 - mean_squared_error: 5.2566e-04 - val_loss: 3.1376e-04 - val_mean_squared_error: 6.2779e-04
Epoch 42/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.6123e-04 - mean_squared_error: 5.2325e-04 - val_loss: 3.3937e-04 - val_mean_squared_error: 6.7904e-04
Epoch 43/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.5928e-04 - mean_squared_error: 5.1934e-04 - val_loss: 2.8976e-04 - val_mean_squared_error: 5.7976e-04
Epoch 44/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.5798e-04 - mean_squared_error: 5.1672e-04 - val_loss: 3.0530e-04 - val_mean_squared_error: 6.1085e-04
Epoch 45/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.5598e-04 - mean_squared_error: 5.1272e-04 - val_loss: 3.0278e-04 - val_mean_squared_error: 6.0582e-04
Epoch 46/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.5508e-04 - mean_squared_error: 5.1092e-04 - val_loss: 2.9635e-04 - val_mean_squared_error: 5.9295e-04
Epoch 47/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.5291e-04 - mean_squared_error: 5.0656e-04 - val_loss: 2.9854e-04 - val_mean_squared_error: 5.9732e-04
Epoch 48/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.5185e-04 - mean_squared_error: 5.0445e-04 - val_loss: 3.4860e-04 - val_mean_squared_error: 6.9751e-04
Epoch 49/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.5085e-04 - mean_squared_error: 5.0244e-04 - val_loss: 3.1082e-04 - val_mean_squared_error: 6.2190e-04
Epoch 50/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.4887e-04 - mean_squared_error: 4.9847e-04 - val_loss: 2.9364e-04 - val_mean_squared_error: 5.8751e-04
Epoch 51/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.4780e-04 - mean_squared_error: 4.9633e-04 - val_loss: 2.9008e-04 - val_mean_squared_error: 5.8039e-04
Epoch 52/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.4674e-04 - mean_squared_error: 4.9421e-04 - val_loss: 3.1582e-04 - val_mean_squared_error: 6.3190e-04
Epoch 53/100
1199/1199 [==============================] - 6s 5ms/step - loss: 2.4565e-04 - mean_squared_error: 4.9202e-04 - val_loss: 3.2495e-04 - val_mean_squared_error: 6.5018e-04

Predictions

pred_test = model.predict(X_test, batch_size = batch_size) 
pred_test = (pred_test * scale_history + center_history) ** 2 #denormalization
pred_test

104/104 [==============================] - 1s 2ms/step

array([[20.93589 ],
       [21.09217 ],
       [21.305178],
       ...,
       [22.345613],
       [21.961597],
       [21.569155]], dtype=float32)

final_df = initial_df[(13000+n_inputs):(13000+n_inputs+pred_test.shape[0])].copy()
final_df["predictions"] = pred_test
final_df

	Day	Close	predictions
13010	2013-09-17	21.222	20.935890
13011	2013-09-18	21.580	21.092171
13012	2013-09-19	21.397	21.305178
13013	2013-09-20	21.002	21.412287
13014	2013-09-23	21.240	21.381840
...	...	...	...
14045	2017-10-25	21.500	23.026583
14046	2017-10-26	21.320	22.664095
14047	2017-10-27	20.790	22.345613
14048	2017-10-30	20.410	21.961597
14049	2017-10-31	20.160	21.569155

1040 rows × 3 columns

Plot predictions vs actual

fig = px.line(final_df, x = "Day", y = ["Close", "predictions"], title = "GE Stock")
fig.update_layout(yaxis_title="Close Value")
fig